Exam guide vs today: the full list
One section for each exam: the differences a comparison of its exam guide with Anthropic's current documentation has found, and what is true today.
Foundations (CCAR-F)
The exam guide (version 1.0, July 2026) and Anthropic's current documentation differ on these points. Learn what the guide tests, and know what's true today. ClaudeArena's questions follow the guide where it's clear, and their explanations mention the current behaviour.
| Topic | Often taught | Today |
|---|---|---|
| Forcing a tool | tool_choice any / named tool | Rejected by Opus 5.5, Sonnet 5.5, Fable 5.1, Mythos 5.1; use auto + strict tool use or structured outputs. Haiku 5.5 still accepts it, but a forced call skips thinking |
| Stop reasons | Six values | Seven, adding model_context_window_exceeded |
| Hook events | Nine | More than thirty; the nine are the core |
| Permission modes | Four | Six, adding auto and dontAsk. default is shown as Manual, and from v2.1.283 auto is the starting mode in the terminal and VS Code |
| Which files loaded? | /memory | /context; /memory lists and edits |
| First positional argument | $1 | $0 (0-based) |
allowed-tools | Restricts tools | Pre-approves tools; disallowed-tools restricts |
| Custom commands | .claude/commands/ | Merged into skills (commands still work) |
| Grep and Glob | Always available | Not in the default set on macOS, Linux and WSL |
| MCP sampling and roots | Current features | Deprecated in spec 2026-07-28 |
| MCP transports | stdio, SSE, Streamable HTTP | stdio and Streamable HTTP; SSE deprecated |
| Thinking | budget_tokens | Adaptive thinking with an effort setting on the newest models |
| Retries | Only 429 and 529 | Also 5xx server errors; SDKs retry twice by default |
| Subagent tool | Task | Agent (renamed in Claude Code v2.1.63); Task now only starts the names of the task-list tools, such as TaskCreate and TaskStop |
| Guaranteed JSON | Tool use with a schema | Only with strict: true or structured outputs; plain tool use can still return a wrong type or miss a field |
| Subagents spawning subagents | Not possible | Possible, up to 3 layers below the main agent by default (CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH) |
| When subagents run | Calls in one turn run in parallel, calls in separate turns in sequence | Subagents run in the background by default, so calls from separate turns can overlap; Claude waits only when it needs the result |
claude -p on untrusted code | Loads the project's configuration like an interactive session | Still does, with no trust dialog, so hooks and MCP servers in a checkout (a fork's PR, say) run; use --bare or --setting-sources user |
| Subagent context | A subagent starts with only the prompt it's given | Still true of a named subagent, but a fork (subagent_type: "fork") inherits the whole conversation, and from Claude Code v2.1.302 Claude starts forks on its own by default in claude -p, the Agent SDK, the desktop app and VS Code |
| Plan mode | No edits until the plan is accepted | Not enforced in an interactive terminal session where bypass permissions is available, such as one started with --dangerously-skip-permissions; still enforced in claude -p, the Agent SDK and VS Code |
| Prefilling the answer | Start the reply with { | Not supported on Claude 4.6 and later (the request returns a 400 error) |
Professional (CCAR-P)
Compared on 30 September 2026: the Professional exam guide (version 1.0, effective July 2026) was re-read, and this guide's Professional pages were checked against Anthropic's documentation. No difference between what the exam guide teaches and today's documentation has been recorded.